In LightFM does the training set need to be the same size as the test set

Question

I'm working on a pure collaborative filtering model in LightFM using a simple user-item matrix.

When I have a training matrix that has more rows than the test matrix, I encounter this error: "Incorrect number of features in user_features"

It seems this is caused by line 789 in lightfm.py:

if not user_features.shape[1] == self.user_embeddings.shape[0]:
    raise ValueError('Incorrect number of features in user_features')

Investigating this - it looks like it is being caused by line 778 in lightfm.py:

n_users, n_items = test_interactions.shape

Which suggests the size of the test dataset should match the train. And making them the same size does fix my error.

How can I persuade LightFM to let me have train and test datasets of different sizes?

Answer 1

I would change the wording. Often in ML, the rows are the samples and in this interpretation your single row-sample is a tuple or something similar (user, item, value). The number of rows in test and train surely has no limitation! But there is one in terms of dimensions. So only in terms of dimensions (which users and which items are observed) those datasets need a 1:1 correspondence. That's maybe a strong assumption, but you can imagine the problem predicting user 11, where there was no user 11 in training (there is no latent-vector which was build).